Research Of Punctate White Matter Lesion Segmentation In Neonates Based On Deep Convolutional Neural Networks

Punctate white matter lesion is a common brain disease.The semantic segmentation of this lesion using efficient automated algorithm in MRI images can assist doctors to analyze it,and then accelerate the diagnosis speed of the disease,so as to reduce the development risk of the newborn in the first time and improve the prognosis.Existing methods often require multiple sequences of MRI images when segmenting the lesion.There are also some complex preprocessing steps in these methods,and their performance is not good enough.Therefore,three efficient semantic segmentation deep learning neural networks are gradually proposed in this thesis based on the characteristics of the punctate white matter lesion.First of all,we proposes a residual U-Net based semantic segmentation model to segment the lesion.This model is mainly composed of residual U-Net,image block segmentation algorithm,focal loss,and fully-connected conditional random field.The residual U-Net can automatically extract the feature information in the MRI image,which avoids the disadvantage of manually setting the feature form based on artificial assumptions.The focus loss can balance the loss of foreground and background and optimize the training process of the model.The image block segmentation algorithm can block the input image into the model to alleviate the imbalance of the number of positive and negative samples caused by the small target area.Fully-connected conditional random fields can optimise prediction probability maps of the lesions based on the pixel feature information in the original image.The performance of the model exceeds the existing related algorithms.Secondly,we proposes a semantic segmentation neural network based on refined R-CNN to segment the disease.The network constructed in this thesis is based on a general two-stage semantic segmentation model.In this model,heuristic candidate region proposal network as well as lightweight segmentation networks are proposed.The heuristic region proposal network can add high-dimensional semantic information around the extracted candidate region,which makes the subsequent classification and positioning network and the segmentation network better process the lesion area.The lightweight segmentation network outperforms the general segmentation network model with fewer parameters when segmenting the lesion area.The results of comparison experiments and ablation experiments show that each module we proposed can improve the performance of the whole model, making the performance of the final model better than other related algorithms.Finally,we proposes a punctate white matter lesion semantic segmentation network based on neighbouring layer feature fusion CNN.The model can assist the segmentation process of the current image slice by using relevant feature information in adjacent image slices,so that the model can perform better in distinguishing lesions from interference.In addition,a self-balanced focus loss function is proposed based on focus loss.The loss function can automatically adjust the weight of each sample loss according to the situation in training,so that the training effect of the model and the performance of lesion segmentation are improved.All three models proposed in this thesis use single-moda data of T1 sequence in the whole process of segmenting punctate white matter lesion in neonates.The preprocessing of the data is also very simple.With the development of experiments and a deeper understanding of the lesion,the performance of these three methods increases sequentially.After experimental verification,the models proposed in this thesis can better segment the lesion area of punctate white matter lesion in neonates.